improve cat design / performance

This used to make a lot of references to design issues with the
SparseArrays package (#2326 /
#20815), which result in a
non-sensical dispatch arrangement, and contribute to a slow loading
experience do to the nonsense Unions that must be checked by subtyping.

Author: vtjnash

base/abstractarray.jl

@@ -1982,19 +1982,18 @@ julia> cat(1, [2], [3;;]; dims=Val(2))
 @inline cat(A...; dims) = _cat(dims, A...)
 # `@constprop :aggressive` allows `catdims` to be propagated as constant improving return type inference
 @constprop :aggressive _cat(catdims, A::AbstractArray{T}...) where {T} = _cat_t(catdims, T, A...)
+const ArrayConcatGroup = Union{Number, AbstractArray}
 
 # The specializations for 1 and 2 inputs are important
 # especially when running with --inline=no, see #11158
-# The specializations for Union{AbstractVecOrMat,Number} are necessary
-# to have more specialized methods here than in LinearAlgebra/uniformscaling.jl
 vcat(A::AbstractArray) = cat(A; dims=Val(1))
 vcat(A::AbstractArray, B::AbstractArray) = cat(A, B; dims=Val(1))
 vcat(A::AbstractArray...) = cat(A...; dims=Val(1))
-vcat(A::Union{AbstractVecOrMat,Number}...) = cat(A...; dims=Val(1))
+vcat(A::Union{AbstractArray,Number}...) = cat(A...; dims=Val(1))
 hcat(A::AbstractArray) = cat(A; dims=Val(2))
 hcat(A::AbstractArray, B::AbstractArray) = cat(A, B; dims=Val(2))
 hcat(A::AbstractArray...) = cat(A...; dims=Val(2))
-hcat(A::Union{AbstractVecOrMat,Number}...) = cat(A...; dims=Val(2))
+hcat(A::Union{AbstractArray,Number}...) = cat(A...; dims=Val(2))
 
 typed_vcat(T::Type, A::AbstractArray) = _cat_t(Val(1), T, A)
 typed_vcat(T::Type, A::AbstractArray, B::AbstractArray) = _cat_t(Val(1), T, A, B)
@@ -2144,8 +2143,6 @@ end
 
 hvcat(rows::Tuple{Vararg{Int}}, xs::Number...) = typed_hvcat(promote_typeof(xs...), rows, xs...)
 hvcat(rows::Tuple{Vararg{Int}}, xs...) = typed_hvcat(promote_eltypeof(xs...), rows, xs...)
-# the following method is needed to provide a more specific one compared to LinearAlgebra/uniformscaling.jl
-hvcat(rows::Tuple{Vararg{Int}}, xs::Union{AbstractVecOrMat,Number}...) = typed_hvcat(promote_eltypeof(xs...), rows, xs...)
 
 function typed_hvcat(::Type{T}, rows::Tuple{Vararg{Int}}, xs::Number...) where T
     nr = length(rows)

base/array.jl

@@ -2001,18 +2001,6 @@ function vcat(arrays::Vector{T}...) where T
 end
 vcat(A::Vector...) = cat(A...; dims=Val(1)) # more special than SparseArrays's vcat
 
-# disambiguation with LinAlg/special.jl
-# Union{Number,Vector,Matrix} is for LinearAlgebra._DenseConcatGroup
-# VecOrMat{T} is for LinearAlgebra._TypedDenseConcatGroup
-hcat(A::Union{Number,Vector,Matrix}...) = cat(A...; dims=Val(2))
-hcat(A::VecOrMat{T}...) where {T} = typed_hcat(T, A...)
-vcat(A::Union{Number,Vector,Matrix}...) = cat(A...; dims=Val(1))
-vcat(A::VecOrMat{T}...) where {T} = typed_vcat(T, A...)
-hvcat(rows::Tuple{Vararg{Int}}, xs::Union{Number,Vector,Matrix}...) =
-    typed_hvcat(promote_eltypeof(xs...), rows, xs...)
-hvcat(rows::Tuple{Vararg{Int}}, xs::VecOrMat{T}...) where {T} =
-    typed_hvcat(T, rows, xs...)
-
 _cat(n::Integer, x::Integer...) = reshape([x...], (ntuple(Returns(1), n-1)..., length(x)))
 
 ## find ##

deps/checksums/SparseArrays-42a02aed4e8499314441ae09f45adc0505ee43b9.tar.gz/md5

@@ -0,0 +1 @@
+b3083759e8ab4ade5c5f9b4c8e52ff2e

deps/checksums/SparseArrays-42a02aed4e8499314441ae09f45adc0505ee43b9.tar.gz/sha512

@@ -0,0 +1 @@
+9334dcd491853cd02505a7536012e2734ac301b9eed46d893d822780c7ae13c337c43f29caf0cb4baa41bc72de049a24c7e0d7a731b91d87730df47aeced1ede

deps/checksums/SparseArrays-8affe9e499379616e33fc60a24bb31500e8423d7.tar.gz/md5

@@ -847,6 +847,9 @@ static jl_array_t *jl_verify_edges(jl_array_t *targets, size_t minworld)
         jl_value_t *expected = jl_array_ptr_ref(targets, i * 3 + 2);
         size_t min_valid = 0;
         size_t max_valid = ~(size_t)0;
+        uint64_t t0 = uv_hrtime() - jl_gc_total_hrtime();
+        uint64_t b0; jl_gc_get_total_bytes(&b0);
+        //if(1);else
         if (invokesig) {
             assert(callee && "unsupported edge");
             jl_methtable_t *mt = jl_method_get_table(((jl_method_instance_t*)callee)->def.method);
@@ -917,9 +920,16 @@ static jl_array_t *jl_verify_edges(jl_array_t *targets, size_t minworld)
             jl_array_ptr_1d_push(_jl_debug_method_invalidation, loctag);
             jl_array_ptr_1d_push(_jl_debug_method_invalidation, matches);
         }
+        //uint64_t t1 = uv_hrtime() - jl_gc_total_hrtime();
+        //uint64_t b1; jl_gc_get_total_bytes(&b1);
+        //ios_printf(ios_stdout, "\n%u\t%u\t%d\t", ((unsigned)(t1 - t0))/1000u, ((unsigned)(b1 - b0))/1024u, max_valid == ~(size_t)0);
+        //jl_static_show((JL_STREAM*)ios_stdout, (jl_value_t*)invokesig);
+        //ios_printf(ios_stdout, "\t");
+        //jl_static_show((JL_STREAM*)ios_stdout, (jl_value_t*)callee);
+        ////
         //jl_static_show((JL_STREAM*)ios_stderr, (jl_value_t*)invokesig);
         //jl_static_show((JL_STREAM*)ios_stderr, (jl_value_t*)callee);
-        //ios_puts(valid ? "valid\n" : "INVALID\n", ios_stderr);
+        //ios_puts(max_valid == ~(size_t)0 ? "valid\n" : "INVALID\n", ios_stderr);
     }
     JL_GC_POP();
     return maxvalids;
@@ -1066,6 +1076,7 @@ static void jl_insert_backedges(jl_array_t *edges, jl_array_t *ext_targets, jl_a
     // determine which CodeInstance objects are still valid in our image
     jl_array_t *valids = jl_verify_edges(ext_targets, minworld);
     JL_GC_PUSH1(&valids);
+
     valids = jl_verify_methods(edges, valids); // consumes edges valids, initializes methods valids
     jl_verify_graph(edges, valids); // propagates methods valids for each edge
     size_t i, l;

deps/checksums/SparseArrays-8affe9e499379616e33fc60a24bb31500e8423d7.tar.gz/sha512

@@ -323,26 +323,12 @@ end
 ==(B::SymTridiagonal, A::Bidiagonal) = A == B
 
 # concatenation
-const _SpecialArrays = Union{Diagonal, Bidiagonal, Tridiagonal, SymTridiagonal}
-const _Symmetric_DenseArrays{T,A<:Matrix} = Symmetric{T,A}
-const _Hermitian_DenseArrays{T,A<:Matrix} = Hermitian{T,A}
-const _Triangular_DenseArrays{T,A<:Matrix} = AbstractTriangular{T,A}
-const _Annotated_DenseArrays = Union{_SpecialArrays, _Triangular_DenseArrays, _Symmetric_DenseArrays, _Hermitian_DenseArrays}
-const _Annotated_Typed_DenseArrays{T} = Union{_Triangular_DenseArrays{T}, _Symmetric_DenseArrays{T}, _Hermitian_DenseArrays{T}}
-const _DenseConcatGroup = Union{Number, Vector, Adjoint{<:Any,<:Vector}, Transpose{<:Any,<:Vector}, Matrix, _Annotated_DenseArrays}
-const _TypedDenseConcatGroup{T} = Union{Vector{T}, Adjoint{T,Vector{T}}, Transpose{T,Vector{T}}, Matrix{T}, _Annotated_Typed_DenseArrays{T}}
-
-promote_to_array_type(::Tuple{Vararg{Union{_DenseConcatGroup,UniformScaling}}}) = Matrix
-
-Base._cat(dims, xs::_DenseConcatGroup...) = Base._cat_t(dims, promote_eltype(xs...), xs...)
-vcat(A::_DenseConcatGroup...) = Base.typed_vcat(promote_eltype(A...), A...)
-hcat(A::_DenseConcatGroup...) = Base.typed_hcat(promote_eltype(A...), A...)
-hvcat(rows::Tuple{Vararg{Int}}, xs::_DenseConcatGroup...) = Base.typed_hvcat(promote_eltype(xs...), rows, xs...)
-# For performance, specially handle the case where the matrices/vectors have homogeneous eltype
-Base._cat(dims, xs::_TypedDenseConcatGroup{T}...) where {T} = Base._cat_t(dims, T, xs...)
-vcat(A::_TypedDenseConcatGroup{T}...) where {T} = Base.typed_vcat(T, A...)
-hcat(A::_TypedDenseConcatGroup{T}...) where {T} = Base.typed_hcat(T, A...)
-hvcat(rows::Tuple{Vararg{Int}}, xs::_TypedDenseConcatGroup{T}...) where {T} = Base.typed_hvcat(T, rows, xs...)
+
+# TODO: remove these deprecations (used by SparseArrays in the past)
+const _DenseConcatGroup = Union{}
+const _SpecialArrays = Union{}
+
+promote_to_array_type(::Tuple) = Matrix
 
 # factorizations
 function cholesky(S::RealHermSymComplexHerm{<:Real,<:SymTridiagonal}, ::NoPivot = NoPivot(); check::Bool = true)

src/staticdata_utils.c

@@ -417,17 +417,16 @@ promote_to_arrays(n,k, ::Type{T}, A, B, C) where {T} =
     (promote_to_arrays_(n[k], T, A), promote_to_arrays_(n[k+1], T, B), promote_to_arrays_(n[k+2], T, C))
 promote_to_arrays(n,k, ::Type{T}, A, B, Cs...) where {T} =
     (promote_to_arrays_(n[k], T, A), promote_to_arrays_(n[k+1], T, B), promote_to_arrays(n,k+2, T, Cs...)...)
-promote_to_array_type(A::Tuple{Vararg{Union{AbstractVecOrMat,UniformScaling,Number}}}) = Matrix
 
 _us2number(A) = A
 _us2number(J::UniformScaling) = J.λ
 
 for (f, _f, dim, name) in ((:hcat, :_hcat, 1, "rows"), (:vcat, :_vcat, 2, "cols"))
     @eval begin
-        @inline $f(A::Union{AbstractVecOrMat,UniformScaling}...) = $_f(A...)
+        @inline $f(A::Union{AbstractArray,UniformScaling}...) = $_f(A...)
         # if there's a Number present, J::UniformScaling must be 1x1-dimensional
-        @inline $f(A::Union{AbstractVecOrMat,UniformScaling,Number}...) = $f(map(_us2number, A)...)
-        function $_f(A::Union{AbstractVecOrMat,UniformScaling,Number}...; array_type = promote_to_array_type(A))
+        @inline $f(A::Union{AbstractArray,UniformScaling,Number}...) = $f(map(_us2number, A)...)
+        function $_f(A::Union{AbstractArray,UniformScaling,Number}...; array_type = promote_to_array_type(A))
             n = -1
             for a in A
                 if !isa(a, UniformScaling)
@@ -445,9 +444,9 @@ for (f, _f, dim, name) in ((:hcat, :_hcat, 1, "rows"), (:vcat, :_vcat, 2, "cols"
     end
 end
 
-hvcat(rows::Tuple{Vararg{Int}}, A::Union{AbstractVecOrMat,UniformScaling}...) = _hvcat(rows, A...)
-hvcat(rows::Tuple{Vararg{Int}}, A::Union{AbstractVecOrMat,UniformScaling,Number}...) = _hvcat(rows, A...)
-function _hvcat(rows::Tuple{Vararg{Int}}, A::Union{AbstractVecOrMat,UniformScaling,Number}...; array_type = promote_to_array_type(A))
+hvcat(rows::Tuple{Vararg{Int}}, A::Union{AbstractArray,UniformScaling}...) = _hvcat(rows, A...)
+hvcat(rows::Tuple{Vararg{Int}}, A::Union{AbstractArray,UniformScaling,Number}...) = _hvcat(rows, A...)
+function _hvcat(rows::Tuple{Vararg{Int}}, A::Union{AbstractArray,UniformScaling,Number}...; array_type = promote_to_array_type(A))
     require_one_based_indexing(A...)
     nr = length(rows)
     sum(rows) == length(A) || throw(ArgumentError("mismatch between row sizes and number of arguments"))

stdlib/LinearAlgebra/src/special.jl

@@ -1,4 +1,4 @@
-SPARSEARRAYS_BRANCH = main
-SPARSEARRAYS_SHA1 = 8affe9e499379616e33fc60a24bb31500e8423d7
+SPARSEARRAYS_BRANCH = jn/better-cats
+SPARSEARRAYS_SHA1 = 42a02aed4e8499314441ae09f45adc0505ee43b9
 SPARSEARRAYS_GIT_URL := https://github.com/JuliaSparse/SparseArrays.jl.git
 SPARSEARRAYS_TAR_URL = https://api.github.com/repos/JuliaSparse/SparseArrays.jl/tarball/$1